File size: 14,461 Bytes
546a9ba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.

import os
import re
import shutil
from string import ascii_uppercase
from tqdm.auto import tqdm
from model.third_party.HMNet.Evaluation.OldROUGEEval import rouge
from model.third_party.HMNet.ThirdParty.ROUGE import pyrouge
from shutil import copyfile
from mpi4py import MPI
import torch
import logging
import json


def write_json_res(
    output_file, tokenizers, x_ids, y_ids, x_tokens, y_tokens, predictions, gts
):
    data = []

    # for x_id, y_id, x_token, y_token, preds, gt in zip(x_ids, y_ids, x_tokens, y_tokens, predictions, gts):
    # x_id = tokenizers[0].decode(x_id, skip_special_tokens=False) if x_id.dim() == 1 else tokenizers[0].convert_tokens_to_string(x_token)
    # y_id = tokenizers[1].decode(y_id, skip_special_tokens=False) if y_id.dim() == 1 else tokenizers[1].convert_tokens_to_string(y_token)
    for x_token, y_token, preds, gt in zip(x_tokens, y_tokens, predictions, gts):
        data.append(
            {
                # 'x_ids': x_id,
                # 'y_ids': y_id,
                "x_tokens": x_token if isinstance(x_token, str) else " ".join(x_token),
                "y_tokens": y_token if isinstance(y_token, str) else " ".join(y_token),
                "predictions": preds,
                "gt": gt,
            }
        )

    json.dump(data, output_file, indent=4, ensure_ascii=False)


logger = logging.getLogger(__name__)

"""
This code can only be run within docker "rouge", because of the usage of rouge-perl
"""


"""" In ROUGE parlance, your summaries are ‘system’ summaries and the gold standard summaries are ‘model’ summaries.
The summaries should be in separate folders, whose paths are set with the system_dir and model_dir variables.
All summaries should contain one sentence per line."""


class ROUGEEval:
    """
    Wrapper class for pyrouge.
    Compute ROUGE given predictions and references for summarization evaluation.
    """

    def __init__(self, run_dir, save_dir, opt):
        self.run_dir = run_dir
        self.save_dir = save_dir
        self.opt = opt

        # use relative path to make it work on Philly
        self.pyrouge_dir = os.path.join(
            os.path.dirname(__file__), "../ThirdParty/ROUGE/ROUGE-1.5.5/"
        )

        self.eval_batches_num = self.opt.get("EVAL_BATCHES_NUM", float("Inf"))
        self.best_score = -float("Inf")
        self.best_res = {}

    def reset_best_score(self, set_high=False):
        if set_high:
            self.best_score = float("Inf")
        else:
            self.best_score = -float("Inf")

    def make_html_safe(self, s):
        s = s.replace("<", "&lt;")
        s = s.replace(">", "&gt;")
        return s

    def print_to_rouge_dir(
        self, summaries, dir, suffix, split_chars, special_char_dict=None
    ):
        for idx, summary in enumerate(summaries):
            fname = os.path.join(dir, "%06d_%s.txt" % (idx, suffix))
            with open(fname, "wb") as f:
                sents = re.split(r"(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s", summary)
                for i, sent in enumerate(sents):
                    if split_chars:
                        # sent = re.sub(r'([\u4e00-\u9fff])', r' \1 ', sent)
                        for x in re.finditer(r"([\u4e00-\u9fff])", sent):
                            if not x.group(1) in special_char_dict:
                                special_char_dict[x.group(1)] = len(special_char_dict)
                            sent = sent.replace(
                                x.group(1), " {} ".format(special_char_dict[x.group(1)])
                            )
                    if i == len(sents) - 1:
                        to_print = sent.encode("utf-8")
                    else:
                        to_print = sent.encode("utf-8") + "\n".encode("utf-8")
                    f.write(to_print)

    def print_to_rouge_dir_gt(self, summaries, dir, suffix, split_chars):
        if split_chars:
            char_dict = {}

        for idx, summary in enumerate(summaries):
            for ref_idx, sub_summary in enumerate(summary.split(" ||| ")):
                fname = os.path.join(
                    dir, "%s.%06d_%s.txt" % (ascii_uppercase[ref_idx], idx, suffix)
                )
                with open(fname, "wb") as f:
                    sents = re.split(
                        r"(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s", sub_summary
                    )
                    for i, sent in enumerate(sents):
                        if split_chars:
                            for x in re.finditer(r"([\u4e00-\u9fff])", sent):
                                if not x.group(1) in char_dict:
                                    char_dict[x.group(1)] = len(char_dict)
                                sent = sent.replace(
                                    x.group(1), " {} ".format(char_dict[x.group(1)])
                                )

                        if i == len(sents) - 1:
                            to_print = sent.encode("utf-8")
                        else:
                            to_print = sent.encode("utf-8") + "\n".encode("utf-8")
                        f.write(to_print)

        if split_chars:
            return char_dict

    # def filter_empty(self, predictions, groundtruths):
    #     new_predicitons = []
    #     new_groundtruths = []
    #
    #     for pred, gt in zip(predictions, groundtruths):
    #         if len(gt) == 0:
    #             continue
    #         new_groundtruths.append(gt)
    #         if len(pred) == 0:
    #             new_predicitons.append('<ept>')
    #         else:
    #             new_predicitons.append(pred)
    #     return new_predicitons, new_groundtruths

    def _convert_tokens_to_string(self, tokenizer, tokens):
        if "EVAL_TOKENIZED" in self.opt:
            tokens = [t for t in tokens if t not in tokenizer.all_special_tokens]
        if "EVAL_LOWERCASE" in self.opt:
            tokens = [t.lower() for t in tokens]
        if "EVAL_TOKENIZED" in self.opt:
            return " ".join(tokens)
        else:
            return tokenizer.decode(
                tokenizer.convert_tokens_to_ids(tokens), skip_special_tokens=True
            )

    def eval_batches(self, module, dev_batches, save_folder, label=""):
        max_sent_len = int(self.opt["MAX_GEN_LENGTH"])

        logger.info(
            "Decoding current model ... \nSaving folder is {}".format(save_folder)
        )

        predictions = []  # prediction of tokens from model
        x_tokens = []  # input tokens
        y_tokens = []  # groundtruths tokens
        x_ids = []  # input token ids
        y_ids = []  # groundtruths token ids
        gts = []  # groundtruths string
        got_better_score = False
        # err = 0
        if not isinstance(module.tokenizer, list):
            encoder_tokenizer = module.tokenizer
            decoder_tokenizer = module.tokenizer
        elif len(module.tokenizer) == 1:
            encoder_tokenizer = module.tokenizer[0]
            decoder_tokenizer = module.tokenizer[0]
        elif len(module.tokenizer) == 2:
            encoder_tokenizer = module.tokenizer[0]
            decoder_tokenizer = module.tokenizer[1]
        else:
            assert False, f"len(module.tokenizer) > 2"

        with torch.no_grad():
            for j, dev_batch in enumerate(dev_batches):
                for b in dev_batch:
                    if torch.is_tensor(dev_batch[b]):
                        dev_batch[b] = dev_batch[b].to(self.opt["device"])

                beam_search_res = module(
                    dev_batch, beam_search=True, max_sent_len=max_sent_len
                )
                pred = [
                    [t[0] for t in x] if len(x) > 0 else [[]] for x in beam_search_res
                ]
                predictions.extend(
                    [
                        [
                            self._convert_tokens_to_string(decoder_tokenizer, tt)
                            for tt in t
                        ]
                        for t in pred
                    ]
                )

                gts.extend(
                    [
                        self._convert_tokens_to_string(decoder_tokenizer, t)
                        for t in dev_batch["decoder_tokens"]
                    ]
                )
                x_tokens.extend(dev_batch["encoder_tokens"])
                y_tokens.extend(dev_batch["decoder_tokens"])

                if ("DEBUG" in self.opt and j >= 10) or j >= self.eval_batches_num:
                    # in debug mode (decode first 10 batches) ortherwise decode first self.eval_batches_num bathes
                    break

        # use MPI to gather results from all processes / GPUs
        # the result of the gather operation is a list of sublists
        # each sublist corresponds to the list created on one of the MPI processes (or GPUs, respectively)
        # we flatten this list into a "simple" list
        assert len(predictions) == len(
            gts
        ), "len(predictions): {0}, len(gts): {1}".format(len(predictions), len(gts))
        comm = MPI.COMM_WORLD
        predictions = comm.gather(predictions, root=0)
        x_tokens = comm.gather(x_tokens, root=0)
        y_tokens = comm.gather(y_tokens, root=0)
        # if GPU numbers are high (>=8), passing x_ids, y_ids to a rank 0 will cause out of memory
        # x_ids = comm.gather(x_ids, root=0)
        # y_ids = comm.gather(y_ids, root=0)
        gts = comm.gather(gts, root=0)
        if self.opt["rank"] == 0:
            # flatten lists
            predictions = [item for sublist in predictions for item in sublist]
            y_tokens = [item for sublist in y_tokens for item in sublist]
            x_tokens = [item for sublist in x_tokens for item in sublist]
            # x_ids = [item for sublist in x_ids for item in sublist]
            # y_ids = [item for sublist in y_ids for item in sublist]
            gts = [item for sublist in gts for item in sublist]
            # import pdb; pdb.set_trace()
            assert (
                len(predictions) == len(y_tokens) == len(x_tokens) == len(gts)
            ), "len(predictions): {0}, len(y_tokens): {1}, len(x_tokens): {2}, len(gts): {3}".format(
                len(predictions), len(y_tokens), len(x_tokens), len(gts)
            )

            # write intermediate results only on rank 0
            if not os.path.isdir(os.path.join(save_folder, "intermediate_results")):
                os.makedirs(os.path.join(save_folder, "intermediate_results"))
            top_1_predictions = [pred[0] for pred in predictions]
            with open(
                os.path.join(
                    save_folder, "intermediate_results", "res_" + label + ".json"
                ),
                "w",
                encoding="utf-8",
            ) as output_file:
                write_json_res(
                    output_file,
                    [encoder_tokenizer, decoder_tokenizer],
                    x_ids,
                    y_ids,
                    x_tokens,
                    y_tokens,
                    predictions,
                    gts,
                )
            try:
                result = self.eval(top_1_predictions, gts)
            except Exception as e:
                logger.exception("ROUGE Eval ERROR")
                result = {}
                score = -float("Inf")
                pass  # this happens when no overlapping between pred and gts
            else:
                rouge_su4 = rouge(top_1_predictions, gts)  # f, prec, recall
                result = {
                    "ROUGE_1": result["rouge_1_f_score"] * 100.0,
                    "ROUGE_1_Prc": result["rouge_1_precision"] * 100.0,
                    "ROUGE_1_Rcl": result["rouge_1_recall"] * 100.0,
                    "ROUGE_2": result["rouge_2_f_score"] * 100.0,
                    "ROUGE_2_Prc": result["rouge_2_precision"] * 100.0,
                    "ROUGE_2_Rcl": result["rouge_2_recall"] * 100.0,
                    "ROUGE_L": result["rouge_l_f_score"] * 100.0,
                    "ROUGE_L_Prc": result["rouge_l_precision"] * 100.0,
                    "ROUGE_L_Rcl": result["rouge_l_recall"] * 100.0,
                    "ROUGE_SU4": rouge_su4["rouge_su4_f_score"] * 100.0,
                }

                score = result["ROUGE_1"]
                if score > self.best_score:
                    copyfile(
                        os.path.join(
                            save_folder,
                            "intermediate_results",
                            "res_" + label + ".json",
                        ),
                        os.path.join(
                            save_folder,
                            "intermediate_results",
                            "res_" + label + ".best.json",
                        ),
                    )
                    self.best_score = score
                    self.best_res = result
                    got_better_score = True

        else:
            result = {}
            score = -float("Inf")
            got_better_score = False

        return result, score, got_better_score

    def eval(self, predictions, groundtruths):
        # predictions, groundtruths = self.filter_empty(predictions, groundtruths)
        predictions = [self.make_html_safe(w) for w in predictions]
        groundtruths = [self.make_html_safe(w) for w in groundtruths]
        pred_dir = os.path.join(self.save_dir, "predictions")
        if os.path.exists(pred_dir):
            shutil.rmtree(pred_dir)
        os.makedirs(pred_dir)

        gt_dir = os.path.join(self.save_dir, "groundtruths")
        if os.path.exists(gt_dir):
            shutil.rmtree(gt_dir)
        os.makedirs(gt_dir)

        special_char_dict = self.print_to_rouge_dir_gt(
            groundtruths, gt_dir, "gt", "SPLIT_CHARS_FOR_EVAL" in self.opt
        )
        self.print_to_rouge_dir(
            predictions,
            pred_dir,
            "pred",
            "SPLIT_CHARS_FOR_EVAL" in self.opt,
            special_char_dict,
        )

        r = pyrouge.Rouge155(self.pyrouge_dir)
        r.system_dir = pred_dir
        r.model_dir = gt_dir
        r.system_filename_pattern = "(\d+)_pred.txt"
        r.model_filename_pattern = "[A-Z].#ID#_gt.txt"
        results = r.output_to_dict(r.convert_and_evaluate())
        return results